Table 7.

Comparison of ESS-MW with different hyperparameter values on House100K

Training stage		Pretext task	Downstream ImageNet classification
α	β	Training loss ↓	Training loss ↓	Test loss ↓	Test accuracy (%) ↑
2	1/60	3.92 ± 0.003	4.62 ± 0.001	4.67 ± 0.004	18.39 ± 0.082∗
2	1/30	3.91 ± 0.004	4.62 ± 0.006	4.69 ± 0.001	18.13 ± 0.067
2	1/120	3.93 ± 0.001	4.63 ± 0.011	4.70 ± 0.016	18.13 ± 0.134
1	1/60	3.97 ± 0.002	4.66 ± 0.006	4.74 ± 0.011	17.69 ± 0.067
4	1/60	3.85 ± 0.002	4.66 ± 0.007	4.70 ± 0.012	17.31 ± 0.102

The best downstream classification result for the datasets is denoted with an asterisk.